1 // SPDX-License-Identifier: GPL-2.0
5 * Builtin report command: Analyze the perf.data input file,
6 * look up and read DSOs and symbol information and display
7 * a histogram of results, along various sorting keys.
11 #include "util/config.h"
13 #include "util/annotate.h"
14 #include "util/color.h"
16 #include <linux/list.h>
17 #include <linux/rbtree.h>
18 #include <linux/err.h>
19 #include <linux/zalloc.h>
21 #include "util/symbol.h"
22 #include "util/map_symbol.h"
23 #include "util/mem-events.h"
24 #include "util/branch.h"
25 #include "util/callchain.h"
26 #include "util/values.h"
29 #include "util/debug.h"
30 #include "util/evlist.h"
31 #include "util/evsel.h"
32 #include "util/evswitch.h"
33 #include "util/header.h"
34 #include "util/session.h"
35 #include "util/srcline.h"
36 #include "util/tool.h"
38 #include <subcmd/parse-options.h>
39 #include <subcmd/exec-cmd.h>
40 #include "util/parse-events.h"
42 #include "util/thread.h"
43 #include "util/sort.h"
44 #include "util/hist.h"
45 #include "util/data.h"
46 #include "arch/common.h"
47 #include "util/time-utils.h"
48 #include "util/auxtrace.h"
49 #include "util/units.h"
50 #include "util/util.h" // perf_tip()
52 #include "ui/progress.h"
53 #include "util/block-info.h"
59 #include <linux/ctype.h>
61 #include <linux/bitmap.h>
62 #include <linux/string.h>
63 #include <linux/stringify.h>
64 #include <linux/time64.h>
65 #include <sys/types.h>
68 #include <linux/mman.h>
71 struct perf_tool tool;
72 struct perf_session *session;
73 struct evswitch evswitch;
74 bool use_tui, use_gtk, use_stdio;
77 bool inverted_callchain;
84 bool nonany_branch_mode;
89 struct perf_read_values show_threads_values;
90 struct annotation_options annotation_opts;
91 const char *pretty_printing_style;
93 const char *symbol_filter_str;
95 struct perf_time_interval *ptime_range;
103 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
104 struct branch_type_stat brtype_stat;
106 bool total_cycles_mode;
107 struct block_report *block_reports;
108 int nr_block_reports;
111 static int report__config(const char *var, const char *value, void *cb)
113 struct report *rep = cb;
115 if (!strcmp(var, "report.group")) {
116 symbol_conf.event_group = perf_config_bool(var, value);
119 if (!strcmp(var, "report.percent-limit")) {
120 double pcnt = strtof(value, NULL);
122 rep->min_percent = pcnt;
123 callchain_param.min_percent = pcnt;
126 if (!strcmp(var, "report.children")) {
127 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
130 if (!strcmp(var, "report.queue-size"))
131 return perf_config_u64(&rep->queue_size, var, value);
133 if (!strcmp(var, "report.sort_order")) {
134 default_sort_order = strdup(value);
141 static int hist_iter__report_callback(struct hist_entry_iter *iter,
142 struct addr_location *al, bool single,
146 struct report *rep = arg;
147 struct hist_entry *he = iter->he;
148 struct evsel *evsel = iter->evsel;
149 struct perf_sample *sample = iter->sample;
151 struct branch_info *bi;
153 if (!ui__has_annotation() && !rep->symbol_ipc)
156 if (sort__mode == SORT_MODE__BRANCH) {
157 bi = he->branch_info;
158 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
162 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
164 } else if (rep->mem_mode) {
166 err = addr_map_symbol__inc_samples(&mi->daddr, sample, evsel);
170 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
172 } else if (symbol_conf.cumulate_callchain) {
174 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
176 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
183 static int hist_iter__branch_callback(struct hist_entry_iter *iter,
184 struct addr_location *al __maybe_unused,
185 bool single __maybe_unused,
188 struct hist_entry *he = iter->he;
189 struct report *rep = arg;
190 struct branch_info *bi = he->branch_info;
191 struct perf_sample *sample = iter->sample;
192 struct evsel *evsel = iter->evsel;
195 branch_type_count(&rep->brtype_stat, &bi->flags,
196 bi->from.addr, bi->to.addr);
198 if (!ui__has_annotation() && !rep->symbol_ipc)
201 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
205 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
211 static void setup_forced_leader(struct report *report,
212 struct evlist *evlist)
214 if (report->group_set)
215 evlist__force_leader(evlist);
218 static int process_feature_event(struct perf_session *session,
219 union perf_event *event)
221 struct report *rep = container_of(session->tool, struct report, tool);
223 if (event->feat.feat_id < HEADER_LAST_FEATURE)
224 return perf_event__process_feature(session, event);
226 if (event->feat.feat_id != HEADER_LAST_FEATURE) {
227 pr_err("failed: wrong feature ID: %" PRI_lu64 "\n",
228 event->feat.feat_id);
230 } else if (rep->header_only) {
235 * (feat_id = HEADER_LAST_FEATURE) is the end marker which
236 * means all features are received, now we can force the
239 setup_forced_leader(rep, session->evlist);
243 static int process_sample_event(struct perf_tool *tool,
244 union perf_event *event,
245 struct perf_sample *sample,
247 struct machine *machine)
249 struct report *rep = container_of(tool, struct report, tool);
250 struct addr_location al;
251 struct hist_entry_iter iter = {
254 .hide_unresolved = symbol_conf.hide_unresolved,
255 .add_entry_cb = hist_iter__report_callback,
259 if (perf_time__ranges_skip_sample(rep->ptime_range, rep->range_num,
264 if (evswitch__discard(&rep->evswitch, evsel))
267 if (machine__resolve(machine, &al, sample) < 0) {
268 pr_debug("problem processing %d event, skipping it.\n",
274 al.thread->lbr_stitch_enable = true;
276 if (symbol_conf.hide_unresolved && al.sym == NULL)
279 if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
282 if (sort__mode == SORT_MODE__BRANCH) {
284 * A non-synthesized event might not have a branch stack if
285 * branch stacks have been synthesized (using itrace options).
287 if (!sample->branch_stack)
290 iter.add_entry_cb = hist_iter__branch_callback;
291 iter.ops = &hist_iter_branch;
292 } else if (rep->mem_mode) {
293 iter.ops = &hist_iter_mem;
294 } else if (symbol_conf.cumulate_callchain) {
295 iter.ops = &hist_iter_cumulative;
297 iter.ops = &hist_iter_normal;
301 al.map->dso->hit = 1;
303 if (ui__has_annotation() || rep->symbol_ipc || rep->total_cycles_mode) {
304 hist__account_cycles(sample->branch_stack, &al, sample,
305 rep->nonany_branch_mode,
309 ret = hist_entry_iter__add(&iter, &al, rep->max_stack, rep);
311 pr_debug("problem adding hist entry, skipping event\n");
313 addr_location__put(&al);
317 static int process_read_event(struct perf_tool *tool,
318 union perf_event *event,
319 struct perf_sample *sample __maybe_unused,
321 struct machine *machine __maybe_unused)
323 struct report *rep = container_of(tool, struct report, tool);
325 if (rep->show_threads) {
326 const char *name = evsel__name(evsel);
327 int err = perf_read_values_add_value(&rep->show_threads_values,
328 event->read.pid, event->read.tid,
340 /* For pipe mode, sample_type is not currently set */
341 static int report__setup_sample_type(struct report *rep)
343 struct perf_session *session = rep->session;
344 u64 sample_type = evlist__combined_sample_type(session->evlist);
345 bool is_pipe = perf_data__is_pipe(session->data);
347 if (session->itrace_synth_opts->callchain ||
348 session->itrace_synth_opts->add_callchain ||
350 perf_header__has_feat(&session->header, HEADER_AUXTRACE) &&
351 !session->itrace_synth_opts->set))
352 sample_type |= PERF_SAMPLE_CALLCHAIN;
354 if (session->itrace_synth_opts->last_branch ||
355 session->itrace_synth_opts->add_last_branch)
356 sample_type |= PERF_SAMPLE_BRANCH_STACK;
358 if (!is_pipe && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
359 if (perf_hpp_list.parent) {
360 ui__error("Selected --sort parent, but no "
361 "callchain data. Did you call "
362 "'perf record' without -g?\n");
365 if (symbol_conf.use_callchain &&
366 !symbol_conf.show_branchflag_count) {
367 ui__error("Selected -g or --branch-history.\n"
368 "But no callchain or branch data.\n"
369 "Did you call 'perf record' without -g or -b?\n");
372 } else if (!callchain_param.enabled &&
373 callchain_param.mode != CHAIN_NONE &&
374 !symbol_conf.use_callchain) {
375 symbol_conf.use_callchain = true;
376 if (callchain_register_param(&callchain_param) < 0) {
377 ui__error("Can't register callchain params.\n");
382 if (symbol_conf.cumulate_callchain) {
383 /* Silently ignore if callchain is missing */
384 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
385 symbol_conf.cumulate_callchain = false;
386 perf_hpp__cancel_cumulate();
390 if (sort__mode == SORT_MODE__BRANCH) {
392 !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
393 ui__error("Selected -b but no branch data. "
394 "Did you call perf record without -b?\n");
399 if (sort__mode == SORT_MODE__MEMORY) {
400 if (!is_pipe && !(sample_type & PERF_SAMPLE_DATA_SRC)) {
401 ui__error("Selected --mem-mode but no mem data. "
402 "Did you call perf record without -d?\n");
407 callchain_param_setup(sample_type);
409 if (rep->stitch_lbr && (callchain_param.record_mode != CALLCHAIN_LBR)) {
410 ui__warning("Can't find LBR callchain. Switch off --stitch-lbr.\n"
411 "Please apply --call-graph lbr when recording.\n");
412 rep->stitch_lbr = false;
415 /* ??? handle more cases than just ANY? */
416 if (!(evlist__combined_branch_type(session->evlist) & PERF_SAMPLE_BRANCH_ANY))
417 rep->nonany_branch_mode = true;
419 #if !defined(HAVE_LIBUNWIND_SUPPORT) && !defined(HAVE_DWARF_SUPPORT)
420 if (dwarf_callchain_users) {
421 ui__warning("Please install libunwind or libdw "
422 "development packages during the perf build.\n");
429 static void sig_handler(int sig __maybe_unused)
434 static size_t hists__fprintf_nr_sample_events(struct hists *hists, struct report *rep,
435 const char *evname, FILE *fp)
439 unsigned long nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
440 u64 nr_events = hists->stats.total_period;
441 struct evsel *evsel = hists_to_evsel(hists);
443 size_t size = sizeof(buf);
444 int socked_id = hists->socket_filter;
449 if (symbol_conf.filter_relative) {
450 nr_samples = hists->stats.nr_non_filtered_samples;
451 nr_events = hists->stats.total_non_filtered_period;
454 if (evsel__is_group_event(evsel)) {
457 evsel__group_desc(evsel, buf, size);
460 for_each_group_member(pos, evsel) {
461 const struct hists *pos_hists = evsel__hists(pos);
463 if (symbol_conf.filter_relative) {
464 nr_samples += pos_hists->stats.nr_non_filtered_samples;
465 nr_events += pos_hists->stats.total_non_filtered_period;
467 nr_samples += pos_hists->stats.nr_events[PERF_RECORD_SAMPLE];
468 nr_events += pos_hists->stats.total_period;
473 nr_samples = convert_unit(nr_samples, &unit);
474 ret = fprintf(fp, "# Samples: %lu%c", nr_samples, unit);
475 if (evname != NULL) {
476 ret += fprintf(fp, " of event%s '%s'",
477 evsel->core.nr_members > 1 ? "s" : "", evname);
481 ret += fprintf(fp, " (time slices: %s)", rep->time_str);
483 if (symbol_conf.show_ref_callgraph && evname && strstr(evname, "call-graph=no")) {
484 ret += fprintf(fp, ", show reference callgraph");
488 ret += fprintf(fp, "\n# Total weight : %" PRIu64, nr_events);
489 ret += fprintf(fp, "\n# Sort order : %s", sort_order ? : default_mem_sort_order);
491 ret += fprintf(fp, "\n# Event count (approx.): %" PRIu64, nr_events);
494 ret += fprintf(fp, "\n# Processor Socket: %d", socked_id);
496 return ret + fprintf(fp, "\n#\n");
499 static int evlist__tui_block_hists_browse(struct evlist *evlist, struct report *rep)
504 evlist__for_each_entry(evlist, pos) {
505 ret = report__browse_block_hists(&rep->block_reports[i++].hist,
506 rep->min_percent, pos,
507 &rep->session->header.env,
508 &rep->annotation_opts);
516 static int evlist__tty_browse_hists(struct evlist *evlist, struct report *rep, const char *help)
522 fprintf(stdout, "#\n# Total Lost Samples: %" PRIu64 "\n#\n",
523 evlist->stats.total_lost_samples);
526 evlist__for_each_entry(evlist, pos) {
527 struct hists *hists = evsel__hists(pos);
528 const char *evname = evsel__name(pos);
530 if (symbol_conf.event_group && !evsel__is_group_leader(pos))
533 hists__fprintf_nr_sample_events(hists, rep, evname, stdout);
535 if (rep->total_cycles_mode) {
536 report__browse_block_hists(&rep->block_reports[i++].hist,
537 rep->min_percent, pos,
542 hists__fprintf(hists, !quiet, 0, 0, rep->min_percent, stdout,
543 !(symbol_conf.use_callchain ||
544 symbol_conf.show_branchflag_count));
545 fprintf(stdout, "\n\n");
549 fprintf(stdout, "#\n# (%s)\n#\n", help);
551 if (rep->show_threads) {
552 bool style = !strcmp(rep->pretty_printing_style, "raw");
553 perf_read_values_display(stdout, &rep->show_threads_values,
555 perf_read_values_destroy(&rep->show_threads_values);
558 if (sort__mode == SORT_MODE__BRANCH)
559 branch_type_stat_display(stdout, &rep->brtype_stat);
564 static void report__warn_kptr_restrict(const struct report *rep)
566 struct map *kernel_map = machine__kernel_map(&rep->session->machines.host);
567 struct kmap *kernel_kmap = kernel_map ? map__kmap(kernel_map) : NULL;
569 if (evlist__exclude_kernel(rep->session->evlist))
572 if (kernel_map == NULL ||
573 (kernel_map->dso->hit &&
574 (kernel_kmap->ref_reloc_sym == NULL ||
575 kernel_kmap->ref_reloc_sym->addr == 0))) {
577 "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
578 "can't be resolved.";
580 if (kernel_map && map__has_symbols(kernel_map)) {
581 desc = "If some relocation was applied (e.g. "
582 "kexec) symbols may be misresolved.";
586 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
587 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
588 "Samples in kernel modules can't be resolved as well.\n\n",
593 static int report__gtk_browse_hists(struct report *rep, const char *help)
595 int (*hist_browser)(struct evlist *evlist, const char *help,
596 struct hist_browser_timer *timer, float min_pcnt);
598 hist_browser = dlsym(perf_gtk_handle, "evlist__gtk_browse_hists");
600 if (hist_browser == NULL) {
601 ui__error("GTK browser not found!\n");
605 return hist_browser(rep->session->evlist, help, NULL, rep->min_percent);
608 static int report__browse_hists(struct report *rep)
611 struct perf_session *session = rep->session;
612 struct evlist *evlist = session->evlist;
613 const char *help = perf_tip(system_path(TIPDIR));
616 /* fallback for people who don't install perf ;-) */
617 help = perf_tip(DOCDIR);
619 help = "Cannot load tips.txt file, please install perf!";
622 switch (use_browser) {
624 if (rep->total_cycles_mode) {
625 ret = evlist__tui_block_hists_browse(evlist, rep);
629 ret = evlist__tui_browse_hists(evlist, help, NULL, rep->min_percent,
630 &session->header.env, true, &rep->annotation_opts);
632 * Usually "ret" is the last pressed key, and we only
633 * care if the key notifies us to switch data file.
635 if (ret != K_SWITCH_INPUT_DATA && ret != K_RELOAD)
639 ret = report__gtk_browse_hists(rep, help);
642 ret = evlist__tty_browse_hists(evlist, rep, help);
649 static int report__collapse_hists(struct report *rep)
651 struct ui_progress prog;
655 ui_progress__init(&prog, rep->nr_entries, "Merging related events...");
657 evlist__for_each_entry(rep->session->evlist, pos) {
658 struct hists *hists = evsel__hists(pos);
661 hists->symbol_filter_str = rep->symbol_filter_str;
663 hists->socket_filter = rep->socket_filter;
665 ret = hists__collapse_resort(hists, &prog);
669 /* Non-group events are considered as leader */
670 if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
671 struct hists *leader_hists = evsel__hists(pos->leader);
673 hists__match(leader_hists, hists);
674 hists__link(leader_hists, hists);
678 ui_progress__finish();
682 static int hists__resort_cb(struct hist_entry *he, void *arg)
684 struct report *rep = arg;
685 struct symbol *sym = he->ms.sym;
687 if (rep->symbol_ipc && sym && !sym->annotate2) {
688 struct evsel *evsel = hists_to_evsel(he->hists);
690 symbol__annotate2(&he->ms, evsel,
691 &annotation__default_options, NULL);
697 static void report__output_resort(struct report *rep)
699 struct ui_progress prog;
702 ui_progress__init(&prog, rep->nr_entries, "Sorting events for output...");
704 evlist__for_each_entry(rep->session->evlist, pos) {
705 evsel__output_resort_cb(pos, &prog, hists__resort_cb, rep);
708 ui_progress__finish();
711 static void stats_setup(struct report *rep)
713 memset(&rep->tool, 0, sizeof(rep->tool));
714 rep->tool.no_warn = true;
717 static int stats_print(struct report *rep)
719 struct perf_session *session = rep->session;
721 perf_session__fprintf_nr_events(session, stdout);
725 static void tasks_setup(struct report *rep)
727 memset(&rep->tool, 0, sizeof(rep->tool));
728 rep->tool.ordered_events = true;
729 if (rep->mmaps_mode) {
730 rep->tool.mmap = perf_event__process_mmap;
731 rep->tool.mmap2 = perf_event__process_mmap2;
733 rep->tool.comm = perf_event__process_comm;
734 rep->tool.exit = perf_event__process_exit;
735 rep->tool.fork = perf_event__process_fork;
736 rep->tool.no_warn = true;
740 struct thread *thread;
741 struct list_head list;
742 struct list_head children;
745 static struct task *tasks_list(struct task *task, struct machine *machine)
747 struct thread *parent_thread, *thread = task->thread;
748 struct task *parent_task;
750 /* Already listed. */
751 if (!list_empty(&task->list))
754 /* Last one in the chain. */
755 if (thread->ppid == -1)
758 parent_thread = machine__find_thread(machine, -1, thread->ppid);
760 return ERR_PTR(-ENOENT);
762 parent_task = thread__priv(parent_thread);
763 list_add_tail(&task->list, &parent_task->children);
764 return tasks_list(parent_task, machine);
767 static size_t maps__fprintf_task(struct maps *maps, int indent, FILE *fp)
772 maps__for_each_entry(maps, map) {
773 printed += fprintf(fp, "%*s %" PRIx64 "-%" PRIx64 " %c%c%c%c %08" PRIx64 " %" PRIu64 " %s\n",
774 indent, "", map->start, map->end,
775 map->prot & PROT_READ ? 'r' : '-',
776 map->prot & PROT_WRITE ? 'w' : '-',
777 map->prot & PROT_EXEC ? 'x' : '-',
778 map->flags & MAP_SHARED ? 's' : 'p',
780 map->dso->id.ino, map->dso->name);
786 static void task__print_level(struct task *task, FILE *fp, int level)
788 struct thread *thread = task->thread;
790 int comm_indent = fprintf(fp, " %8d %8d %8d |%*s",
791 thread->pid_, thread->tid, thread->ppid,
794 fprintf(fp, "%s\n", thread__comm_str(thread));
796 maps__fprintf_task(thread->maps, comm_indent, fp);
798 if (!list_empty(&task->children)) {
799 list_for_each_entry(child, &task->children, list)
800 task__print_level(child, fp, level + 1);
804 static int tasks_print(struct report *rep, FILE *fp)
806 struct perf_session *session = rep->session;
807 struct machine *machine = &session->machines.host;
808 struct task *tasks, *task;
809 unsigned int nr = 0, itask = 0, i;
814 * No locking needed while accessing machine->threads,
815 * because --tasks is single threaded command.
818 /* Count all the threads. */
819 for (i = 0; i < THREADS__TABLE_SIZE; i++)
820 nr += machine->threads[i].nr;
822 tasks = malloc(sizeof(*tasks) * nr);
826 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
827 struct threads *threads = &machine->threads[i];
829 for (nd = rb_first_cached(&threads->entries); nd;
831 task = tasks + itask++;
833 task->thread = rb_entry(nd, struct thread, rb_node);
834 INIT_LIST_HEAD(&task->children);
835 INIT_LIST_HEAD(&task->list);
836 thread__set_priv(task->thread, task);
841 * Iterate every task down to the unprocessed parent
842 * and link all in task children list. Task with no
843 * parent is added into 'list'.
845 for (itask = 0; itask < nr; itask++) {
846 task = tasks + itask;
848 if (!list_empty(&task->list))
851 task = tasks_list(task, machine);
853 pr_err("Error: failed to process tasks\n");
855 return PTR_ERR(task);
859 list_add_tail(&task->list, &list);
862 fprintf(fp, "# %8s %8s %8s %s\n", "pid", "tid", "ppid", "comm");
864 list_for_each_entry(task, &list, list)
865 task__print_level(task, fp, 0);
871 static int __cmd_report(struct report *rep)
874 struct perf_session *session = rep->session;
876 struct perf_data *data = session->data;
878 signal(SIGINT, sig_handler);
881 ret = perf_session__cpu_bitmap(session, rep->cpu_list,
884 ui__error("failed to set cpu bitmap\n");
887 session->itrace_synth_opts->cpu_bitmap = rep->cpu_bitmap;
890 if (rep->show_threads) {
891 ret = perf_read_values_init(&rep->show_threads_values);
896 ret = report__setup_sample_type(rep);
898 /* report__setup_sample_type() already showed error message */
908 ret = perf_session__process_events(session);
910 ui__error("failed to process sample\n");
915 return stats_print(rep);
918 return tasks_print(rep, stdout);
920 report__warn_kptr_restrict(rep);
922 evlist__for_each_entry(session->evlist, pos)
923 rep->nr_entries += evsel__hists(pos)->nr_entries;
925 if (use_browser == 0) {
927 perf_session__fprintf(session, stdout);
930 perf_session__fprintf_dsos(session, stdout);
933 perf_session__fprintf_nr_events(session, stdout);
934 evlist__fprintf_nr_events(session->evlist, stdout);
939 ret = report__collapse_hists(rep);
941 ui__error("failed to process hist entry\n");
949 * recalculate number of entries after collapsing since it
950 * might be changed during the collapse phase.
953 evlist__for_each_entry(session->evlist, pos)
954 rep->nr_entries += evsel__hists(pos)->nr_entries;
956 if (rep->nr_entries == 0) {
957 ui__error("The %s data has no samples!\n", data->path);
961 report__output_resort(rep);
963 if (rep->total_cycles_mode) {
964 int block_hpps[6] = {
965 PERF_HPP_REPORT__BLOCK_TOTAL_CYCLES_PCT,
966 PERF_HPP_REPORT__BLOCK_LBR_CYCLES,
967 PERF_HPP_REPORT__BLOCK_CYCLES_PCT,
968 PERF_HPP_REPORT__BLOCK_AVG_CYCLES,
969 PERF_HPP_REPORT__BLOCK_RANGE,
970 PERF_HPP_REPORT__BLOCK_DSO,
973 rep->block_reports = block_info__create_report(session->evlist,
976 &rep->nr_block_reports);
977 if (!rep->block_reports)
981 return report__browse_hists(rep);
985 report_parse_callchain_opt(const struct option *opt, const char *arg, int unset)
987 struct callchain_param *callchain = opt->value;
989 callchain->enabled = !unset;
994 symbol_conf.use_callchain = false;
995 callchain->mode = CHAIN_NONE;
999 return parse_callchain_report_opt(arg);
1003 parse_time_quantum(const struct option *opt, const char *arg,
1004 int unset __maybe_unused)
1006 unsigned long *time_q = opt->value;
1009 *time_q = strtoul(arg, &end, 0);
1013 pr_err("time quantum cannot be 0");
1016 end = skip_spaces(end);
1019 if (!strcmp(end, "s")) {
1020 *time_q *= NSEC_PER_SEC;
1023 if (!strcmp(end, "ms")) {
1024 *time_q *= NSEC_PER_MSEC;
1027 if (!strcmp(end, "us")) {
1028 *time_q *= NSEC_PER_USEC;
1031 if (!strcmp(end, "ns"))
1034 pr_err("Cannot parse time quantum `%s'\n", arg);
1039 report_parse_ignore_callees_opt(const struct option *opt __maybe_unused,
1040 const char *arg, int unset __maybe_unused)
1043 int err = regcomp(&ignore_callees_regex, arg, REG_EXTENDED);
1046 regerror(err, &ignore_callees_regex, buf, sizeof(buf));
1047 pr_err("Invalid --ignore-callees regex: %s\n%s", arg, buf);
1050 have_ignore_callees = 1;
1057 parse_branch_mode(const struct option *opt,
1058 const char *str __maybe_unused, int unset)
1060 int *branch_mode = opt->value;
1062 *branch_mode = !unset;
1067 parse_percent_limit(const struct option *opt, const char *str,
1068 int unset __maybe_unused)
1070 struct report *rep = opt->value;
1071 double pcnt = strtof(str, NULL);
1073 rep->min_percent = pcnt;
1074 callchain_param.min_percent = pcnt;
1078 static int process_attr(struct perf_tool *tool __maybe_unused,
1079 union perf_event *event,
1080 struct evlist **pevlist)
1085 err = perf_event__process_attr(tool, event, pevlist);
1090 * Check if we need to enable callchains based
1091 * on events sample_type.
1093 sample_type = evlist__combined_sample_type(*pevlist);
1094 callchain_param_setup(sample_type);
1098 int cmd_report(int argc, const char **argv)
1100 struct perf_session *session;
1101 struct itrace_synth_opts itrace_synth_opts = { .set = 0, };
1103 bool has_br_stack = false;
1104 int branch_mode = -1;
1106 bool branch_call_mode = false;
1107 #define CALLCHAIN_DEFAULT_OPT "graph,0.5,caller,function,percent"
1108 static const char report_callchain_help[] = "Display call graph (stack chain/backtrace):\n\n"
1109 CALLCHAIN_REPORT_HELP
1110 "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT;
1111 char callchain_default_opt[] = CALLCHAIN_DEFAULT_OPT;
1112 const char * const report_usage[] = {
1113 "perf report [<options>]",
1116 struct report report = {
1118 .sample = process_sample_event,
1119 .mmap = perf_event__process_mmap,
1120 .mmap2 = perf_event__process_mmap2,
1121 .comm = perf_event__process_comm,
1122 .namespaces = perf_event__process_namespaces,
1123 .cgroup = perf_event__process_cgroup,
1124 .exit = perf_event__process_exit,
1125 .fork = perf_event__process_fork,
1126 .lost = perf_event__process_lost,
1127 .read = process_read_event,
1128 .attr = process_attr,
1129 .tracing_data = perf_event__process_tracing_data,
1130 .build_id = perf_event__process_build_id,
1131 .id_index = perf_event__process_id_index,
1132 .auxtrace_info = perf_event__process_auxtrace_info,
1133 .auxtrace = perf_event__process_auxtrace,
1134 .event_update = perf_event__process_event_update,
1135 .feature = process_feature_event,
1136 .ordered_events = true,
1137 .ordering_requires_timestamps = true,
1139 .max_stack = PERF_MAX_STACK_DEPTH,
1140 .pretty_printing_style = "normal",
1141 .socket_filter = -1,
1142 .annotation_opts = annotation__default_options,
1144 const struct option options[] = {
1145 OPT_STRING('i', "input", &input_name, "file",
1147 OPT_INCR('v', "verbose", &verbose,
1148 "be more verbose (show symbol address, etc)"),
1149 OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any message"),
1150 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1151 "dump raw trace in ASCII"),
1152 OPT_BOOLEAN(0, "stats", &report.stats_mode, "Display event stats"),
1153 OPT_BOOLEAN(0, "tasks", &report.tasks_mode, "Display recorded tasks"),
1154 OPT_BOOLEAN(0, "mmaps", &report.mmaps_mode, "Display recorded tasks memory maps"),
1155 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1156 "file", "vmlinux pathname"),
1157 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1158 "don't load vmlinux even if found"),
1159 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1160 "file", "kallsyms pathname"),
1161 OPT_BOOLEAN('f', "force", &symbol_conf.force, "don't complain, do it"),
1162 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
1163 "load module symbols - WARNING: use only with -k and LIVE kernel"),
1164 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1165 "Show a column with the number of samples"),
1166 OPT_BOOLEAN('T', "threads", &report.show_threads,
1167 "Show per-thread event counters"),
1168 OPT_STRING(0, "pretty", &report.pretty_printing_style, "key",
1169 "pretty printing style key: normal raw"),
1170 OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"),
1171 OPT_BOOLEAN(0, "gtk", &report.use_gtk, "Use the GTK2 interface"),
1172 OPT_BOOLEAN(0, "stdio", &report.use_stdio,
1173 "Use the stdio interface"),
1174 OPT_BOOLEAN(0, "header", &report.header, "Show data header."),
1175 OPT_BOOLEAN(0, "header-only", &report.header_only,
1176 "Show only data header."),
1177 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1178 sort_help("sort by key(s):")),
1179 OPT_STRING('F', "fields", &field_order, "key[,keys...]",
1180 sort_help("output field(s): overhead period sample ")),
1181 OPT_BOOLEAN(0, "show-cpu-utilization", &symbol_conf.show_cpu_utilization,
1182 "Show sample percentage for different cpu modes"),
1183 OPT_BOOLEAN_FLAG(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
1184 "Show sample percentage for different cpu modes", PARSE_OPT_HIDDEN),
1185 OPT_STRING('p', "parent", &parent_pattern, "regex",
1186 "regex filter to identify parent, see: '--sort parent'"),
1187 OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
1188 "Only display entries with parent-match"),
1189 OPT_CALLBACK_DEFAULT('g', "call-graph", &callchain_param,
1190 "print_type,threshold[,print_limit],order,sort_key[,branch],value",
1191 report_callchain_help, &report_parse_callchain_opt,
1192 callchain_default_opt),
1193 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1194 "Accumulate callchains of children and show total overhead as well. "
1195 "Enabled by default, use --no-children to disable."),
1196 OPT_INTEGER(0, "max-stack", &report.max_stack,
1197 "Set the maximum stack depth when parsing the callchain, "
1198 "anything beyond the specified depth will be ignored. "
1199 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1200 OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
1201 "alias for inverted call graph"),
1202 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1203 "ignore callees of these functions in call graphs",
1204 report_parse_ignore_callees_opt),
1205 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1206 "only consider symbols in these dsos"),
1207 OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1208 "only consider symbols in these comms"),
1209 OPT_STRING(0, "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
1210 "only consider symbols in these pids"),
1211 OPT_STRING(0, "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
1212 "only consider symbols in these tids"),
1213 OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1214 "only consider these symbols"),
1215 OPT_STRING(0, "symbol-filter", &report.symbol_filter_str, "filter",
1216 "only show symbols that (partially) match with this filter"),
1217 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1219 "don't try to adjust column width, use these fixed values"),
1220 OPT_STRING_NOEMPTY('t', "field-separator", &symbol_conf.field_sep, "separator",
1221 "separator for columns, no spaces will be added between "
1222 "columns '.' is reserved."),
1223 OPT_BOOLEAN('U', "hide-unresolved", &symbol_conf.hide_unresolved,
1224 "Only display entries resolved to a symbol"),
1225 OPT_CALLBACK(0, "symfs", NULL, "directory",
1226 "Look for files with symbols relative to this directory",
1227 symbol__config_symfs),
1228 OPT_STRING('C', "cpu", &report.cpu_list, "cpu",
1229 "list of cpus to profile"),
1230 OPT_BOOLEAN('I', "show-info", &report.show_full_info,
1231 "Display extended information about perf.data file"),
1232 OPT_BOOLEAN(0, "source", &report.annotation_opts.annotate_src,
1233 "Interleave source code with assembly code (default)"),
1234 OPT_BOOLEAN(0, "asm-raw", &report.annotation_opts.show_asm_raw,
1235 "Display raw encoding of assembly instructions (default)"),
1236 OPT_STRING('M', "disassembler-style", &report.annotation_opts.disassembler_style, "disassembler style",
1237 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1238 OPT_STRING(0, "prefix", &report.annotation_opts.prefix, "prefix",
1239 "Add prefix to source file path names in programs (with --prefix-strip)"),
1240 OPT_STRING(0, "prefix-strip", &report.annotation_opts.prefix_strip, "N",
1241 "Strip first N entries of source file path name in programs (with --prefix)"),
1242 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1243 "Show a column with the sum of periods"),
1244 OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group, &report.group_set,
1245 "Show event group information together"),
1246 OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1247 "Sort the output by the event at the index n in group. "
1248 "If n is invalid, sort by the first event. "
1249 "WARNING: should be used on grouped events."),
1250 OPT_CALLBACK_NOOPT('b', "branch-stack", &branch_mode, "",
1251 "use branch records for per branch histogram filling",
1253 OPT_BOOLEAN(0, "branch-history", &branch_call_mode,
1254 "add last branch records to call history"),
1255 OPT_STRING(0, "objdump", &report.annotation_opts.objdump_path, "path",
1256 "objdump binary to use for disassembly and annotations"),
1257 OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
1258 "Disable symbol demangling"),
1259 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1260 "Enable kernel symbol demangling"),
1261 OPT_BOOLEAN(0, "mem-mode", &report.mem_mode, "mem access profile"),
1262 OPT_INTEGER(0, "samples", &symbol_conf.res_sample,
1263 "Number of samples to save per histogram entry for individual browsing"),
1264 OPT_CALLBACK(0, "percent-limit", &report, "percent",
1265 "Don't show entries under that percent", parse_percent_limit),
1266 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1267 "how to display percentage of filtered entries", parse_filter_percentage),
1268 OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts",
1269 "Instruction Tracing options\n" ITRACE_HELP,
1270 itrace_parse_synth_opts),
1271 OPT_BOOLEAN(0, "full-source-path", &srcline_full_filename,
1272 "Show full source file name path for source lines"),
1273 OPT_BOOLEAN(0, "show-ref-call-graph", &symbol_conf.show_ref_callgraph,
1274 "Show callgraph from reference event"),
1275 OPT_BOOLEAN(0, "stitch-lbr", &report.stitch_lbr,
1276 "Enable LBR callgraph stitching approach"),
1277 OPT_INTEGER(0, "socket-filter", &report.socket_filter,
1278 "only show processor socket that match with this filter"),
1279 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1280 "Show raw trace event output (do not use print fmt or plugins)"),
1281 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1282 "Show entries in a hierarchy"),
1283 OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
1284 "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
1285 stdio__config_color, "always"),
1286 OPT_STRING(0, "time", &report.time_str, "str",
1287 "Time span of interest (start,stop)"),
1288 OPT_BOOLEAN(0, "inline", &symbol_conf.inline_name,
1289 "Show inline function"),
1290 OPT_CALLBACK(0, "percent-type", &report.annotation_opts, "local-period",
1291 "Set percent type local/global-period/hits",
1292 annotate_parse_percent_type),
1293 OPT_BOOLEAN(0, "ns", &symbol_conf.nanosecs, "Show times in nanosecs"),
1294 OPT_CALLBACK(0, "time-quantum", &symbol_conf.time_quantum, "time (ms|us|ns|s)",
1295 "Set time quantum for time sort key (default 100ms)",
1296 parse_time_quantum),
1297 OPTS_EVSWITCH(&report.evswitch),
1298 OPT_BOOLEAN(0, "total-cycles", &report.total_cycles_mode,
1299 "Sort all blocks by 'Sampled Cycles%'"),
1300 OPT_BOOLEAN(0, "disable-order", &report.disable_order,
1301 "Disable raw trace ordering"),
1304 struct perf_data data = {
1305 .mode = PERF_DATA_MODE_READ,
1307 int ret = hists__init();
1313 ret = perf_config(report__config, &report);
1317 argc = parse_options(argc, argv, options, report_usage, 0);
1320 * Special case: if there's an argument left then assume that
1321 * it's a symbol filter:
1324 usage_with_options(report_usage, options);
1326 report.symbol_filter_str = argv[0];
1329 if (annotate_check_args(&report.annotation_opts) < 0)
1332 if (report.mmaps_mode)
1333 report.tasks_mode = true;
1335 if (dump_trace && report.disable_order)
1336 report.tool.ordered_events = false;
1339 perf_quiet_option();
1341 if (symbol_conf.vmlinux_name &&
1342 access(symbol_conf.vmlinux_name, R_OK)) {
1343 pr_err("Invalid file: %s\n", symbol_conf.vmlinux_name);
1346 if (symbol_conf.kallsyms_name &&
1347 access(symbol_conf.kallsyms_name, R_OK)) {
1348 pr_err("Invalid file: %s\n", symbol_conf.kallsyms_name);
1352 if (report.inverted_callchain)
1353 callchain_param.order = ORDER_CALLER;
1354 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1355 callchain_param.order = ORDER_CALLER;
1357 if ((itrace_synth_opts.callchain || itrace_synth_opts.add_callchain) &&
1358 (int)itrace_synth_opts.callchain_sz > report.max_stack)
1359 report.max_stack = itrace_synth_opts.callchain_sz;
1361 if (!input_name || !strlen(input_name)) {
1362 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1365 input_name = "perf.data";
1368 data.path = input_name;
1369 data.force = symbol_conf.force;
1372 session = perf_session__new(&data, false, &report.tool);
1373 if (IS_ERR(session))
1374 return PTR_ERR(session);
1376 ret = evswitch__init(&report.evswitch, session->evlist, stderr);
1380 if (zstd_init(&(session->zstd_data), 0) < 0)
1381 pr_warning("Decompression initialization failed. Reported data may be incomplete.\n");
1383 if (report.queue_size) {
1384 ordered_events__set_alloc_size(&session->ordered_events,
1388 session->itrace_synth_opts = &itrace_synth_opts;
1390 report.session = session;
1392 has_br_stack = perf_header__has_feat(&session->header,
1393 HEADER_BRANCH_STACK);
1394 if (evlist__combined_sample_type(session->evlist) & PERF_SAMPLE_STACK_USER)
1395 has_br_stack = false;
1397 setup_forced_leader(&report, session->evlist);
1399 if (symbol_conf.group_sort_idx && !session->evlist->nr_groups) {
1400 parse_options_usage(NULL, options, "group-sort-idx", 0);
1405 if (itrace_synth_opts.last_branch || itrace_synth_opts.add_last_branch)
1406 has_br_stack = true;
1408 if (has_br_stack && branch_call_mode)
1409 symbol_conf.show_branchflag_count = true;
1411 memset(&report.brtype_stat, 0, sizeof(struct branch_type_stat));
1414 * Branch mode is a tristate:
1415 * -1 means default, so decide based on the file having branch data.
1416 * 0/1 means the user chose a mode.
1418 if (((branch_mode == -1 && has_br_stack) || branch_mode == 1) &&
1419 !branch_call_mode) {
1420 sort__mode = SORT_MODE__BRANCH;
1421 symbol_conf.cumulate_callchain = false;
1423 if (branch_call_mode) {
1424 callchain_param.key = CCKEY_ADDRESS;
1425 callchain_param.branch_callstack = true;
1426 symbol_conf.use_callchain = true;
1427 callchain_register_param(&callchain_param);
1428 if (sort_order == NULL)
1429 sort_order = "srcline,symbol,dso";
1432 if (report.mem_mode) {
1433 if (sort__mode == SORT_MODE__BRANCH) {
1434 pr_err("branch and mem mode incompatible\n");
1437 sort__mode = SORT_MODE__MEMORY;
1438 symbol_conf.cumulate_callchain = false;
1441 if (symbol_conf.report_hierarchy) {
1442 /* disable incompatible options */
1443 symbol_conf.cumulate_callchain = false;
1446 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1447 parse_options_usage(report_usage, options, "F", 1);
1448 parse_options_usage(NULL, options, "hierarchy", 0);
1452 perf_hpp_list.need_collapse = true;
1455 if (report.use_stdio)
1457 else if (report.use_tui)
1459 else if (report.use_gtk)
1462 /* Force tty output for header output and per-thread stat. */
1463 if (report.header || report.header_only || report.show_threads)
1465 if (report.header || report.header_only)
1466 report.tool.show_feat_hdr = SHOW_FEAT_HEADER;
1467 if (report.show_full_info)
1468 report.tool.show_feat_hdr = SHOW_FEAT_HEADER_FULL_INFO;
1469 if (report.stats_mode || report.tasks_mode)
1471 if (report.stats_mode && report.tasks_mode) {
1472 pr_err("Error: --tasks and --mmaps can't be used together with --stats\n");
1476 if (report.total_cycles_mode) {
1477 if (sort__mode != SORT_MODE__BRANCH)
1478 report.total_cycles_mode = false;
1483 if (strcmp(input_name, "-") != 0)
1484 setup_browser(true);
1488 if (sort_order && strstr(sort_order, "ipc")) {
1489 parse_options_usage(report_usage, options, "s", 1);
1493 if (sort_order && strstr(sort_order, "symbol")) {
1494 if (sort__mode == SORT_MODE__BRANCH) {
1495 snprintf(sort_tmp, sizeof(sort_tmp), "%s,%s",
1496 sort_order, "ipc_lbr");
1497 report.symbol_ipc = true;
1499 snprintf(sort_tmp, sizeof(sort_tmp), "%s,%s",
1500 sort_order, "ipc_null");
1503 sort_order = sort_tmp;
1506 if ((last_key != K_SWITCH_INPUT_DATA && last_key != K_RELOAD) &&
1507 (setup_sorting(session->evlist) < 0)) {
1509 parse_options_usage(report_usage, options, "s", 1);
1511 parse_options_usage(sort_order ? NULL : report_usage,
1516 if ((report.header || report.header_only) && !quiet) {
1517 perf_session__fprintf_info(session, stdout,
1518 report.show_full_info);
1519 if (report.header_only) {
1522 * we need to process first few records
1523 * which contains PERF_RECORD_HEADER_FEATURE.
1525 perf_session__process_events(session);
1530 } else if (use_browser == 0 && !quiet &&
1531 !report.stats_mode && !report.tasks_mode) {
1532 fputs("# To display the perf.data header info, please use --header/--header-only options.\n#\n",
1537 * Only in the TUI browser we are doing integrated annotation,
1538 * so don't allocate extra space that won't be used in the stdio
1541 if (ui__has_annotation() || report.symbol_ipc ||
1542 report.total_cycles_mode) {
1543 ret = symbol__annotation_init();
1547 * For searching by name on the "Browse map details".
1548 * providing it only in verbose mode not to bloat too
1549 * much struct symbol.
1553 * XXX: Need to provide a less kludgy way to ask for
1554 * more space per symbol, the u32 is for the index on
1556 * See symbol__browser_index.
1558 symbol_conf.priv_size += sizeof(u32);
1559 symbol_conf.sort_by_name = true;
1561 annotation_config__init(&report.annotation_opts);
1564 if (symbol__init(&session->header.env) < 0)
1567 if (report.time_str) {
1568 ret = perf_time__parse_for_ranges(report.time_str, session,
1569 &report.ptime_range,
1575 itrace_synth_opts__set_time_range(&itrace_synth_opts,
1580 if (session->tevent.pevent &&
1581 tep_set_function_resolver(session->tevent.pevent,
1582 machine__resolve_kernel_addr,
1583 &session->machines.host) < 0) {
1584 pr_err("%s: failed to set libtraceevent function resolver\n",
1589 sort__setup_elide(stdout);
1591 ret = __cmd_report(&report);
1592 if (ret == K_SWITCH_INPUT_DATA || ret == K_RELOAD) {
1593 perf_session__delete(session);
1594 last_key = K_SWITCH_INPUT_DATA;
1600 if (report.ptime_range) {
1601 itrace_synth_opts__clear_time_range(&itrace_synth_opts);
1602 zfree(&report.ptime_range);
1605 if (report.block_reports) {
1606 block_info__free_report(report.block_reports,
1607 report.nr_block_reports);
1608 report.block_reports = NULL;
1611 zstd_fini(&(session->zstd_data));
1612 perf_session__delete(session);